fix(pyrowave): per-session raw-dmabuf zero-copy capture on the Linux NVIDIA host
ci / web (push) Successful in 52s
apple / swift (push) Successful in 1m13s
ci / docs-site (push) Successful in 1m11s
ci / bench (push) Successful in 5m35s
apple / screenshots (push) Successful in 6m28s
docker / build-push (--build-arg FEDORA_VERSION=44, ci, ci/fedora-rpm.Dockerfile, punktfunk-fedora44-rpm) (push) Successful in 12s
decky / build-publish (push) Successful in 29s
docker / build-push (., web/Dockerfile, punktfunk-web) (push) Successful in 11s
docker / build-push (ci, ci/fedora-rpm.Dockerfile, punktfunk-fedora-rpm) (push) Successful in 10s
docker / build-push (ci, ci/rust-ci-noble.Dockerfile, punktfunk-rust-ci-noble) (push) Successful in 9s
docker / build-push (ci, ci/rust-ci.Dockerfile, punktfunk-rust-ci) (push) Successful in 10s
docker / build-push (docs-site, docs-site/Dockerfile, punktfunk-docs) (push) Successful in 12s
android / android (push) Successful in 13m34s
deb / build-publish (push) Successful in 9m10s
deb / build-publish-host (push) Successful in 9m26s
arch / build-publish (push) Successful in 17m58s
windows-host / package (push) Successful in 16m40s
ci / rust (push) Successful in 29m45s
rpm / build-publish (43, bazzite, punktfunk-fedora-rpm) (push) Successful in 20m12s
rpm / build-publish (44, fedora-44, punktfunk-fedora44-rpm) (push) Successful in 20m8s
docker / deploy-docs (push) Successful in 28s

A PyroWave session on an NVIDIA-auto host was forced onto CPU-RGB capture
(session_plan flipped gpu=false): Mutter blits tiled->LINEAR, we mmap +
de-pad ~30 MB, the encoder re-uploads it - three full-frame CPU touches
per frame at 5120x1440 while an HEVC session on the same box rides the
tiled EGL/CUDA zero-copy. The dmabuf passthrough + Vulkan tiled import
were already validated (8dc5d672) but only reachable via the global
PUNKTFUNK_ENCODER=pyrowave lab policy.

ZeroCopyPolicy gains pyrowave_session (from OutputFormat.pyrowave, i.e.
the negotiated codec): the capturer skips the NVENC-only EGL->CUDA
importer, takes the raw-dmabuf passthrough, and advertises the wavelet
encoder's Vulkan-importable modifiers so Mutter+NVIDIA negotiates tiled
zero-copy. The forced-CPU flip in session_plan is gone.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
enricobuehler
2026-07-19 15:57:12 +02:00
committed by enricobuehler
parent 1d587a259e
commit d2daeacc60
4 changed files with 65 additions and 54 deletions
+23 -16
View File
@@ -255,9 +255,11 @@ fn spawn_pipewire(
want_hdr
};
// Mirror of the thread's `vaapi_passthrough` decision (deterministic from here: on a VAAPI
// backend the EGL→CUDA importer is never built) — kept on the capturer so `next_frame`'s
// negotiation-timeout branch knows a failed negotiation was the LINEAR-dmabuf offer.
let vaapi_dmabuf = zerocopy && !force_shm && policy.backend_is_vaapi;
// backend or a PyroWave session the EGL→CUDA importer is never built) — kept on the capturer
// so `next_frame`'s negotiation-timeout branch knows a failed negotiation was the raw-dmabuf
// passthrough offer.
let vaapi_dmabuf =
zerocopy && !force_shm && (policy.backend_is_vaapi || policy.pyrowave_session);
let join = thread::Builder::new()
.name("punktfunk-pipewire".into())
.spawn(move || {
@@ -1945,16 +1947,18 @@ mod pipewire {
// Build the GPU importer up front — normally the ISOLATED worker process
// (design/zerocopy-worker-isolation.md), so a driver fault on a dying compositor's
// dmabuf kills the worker, not this host. If it fails, log and fall back to the CPU path
// (we simply won't request dmabuf below). Skipped entirely when the encode backend is
// VAAPI: those frames go to the raw-dmabuf passthrough, and building the importer there
// would waste a CUDA probe — or worse, on an NVIDIA box forced to PUNKTFUNK_ENCODER=vaapi,
// succeed and produce CUDA payloads the VAAPI encoder must reject. Also skipped once
// repeated worker deaths latched the import off (a wedged GPU stack must not crash-loop).
// (we simply won't request dmabuf below). Skipped entirely when the frames go to the
// raw-dmabuf passthrough — the encode backend is VAAPI, or the SESSION encodes PyroWave
// (its Vulkan device imports raw dmabufs on any vendor): building the importer there
// would waste a CUDA probe — or worse, succeed and produce CUDA payloads only NVENC can
// consume. Also skipped once repeated worker deaths latched the import off (a wedged GPU
// stack must not crash-loop).
let backend_is_vaapi = policy.backend_is_vaapi;
let raw_passthrough = backend_is_vaapi || policy.pyrowave_session;
// HDR never builds the EGL→CUDA importer: its de-tile blit renders into 8-bit RGBA8,
// which would silently crush the 10-bit depth. The HDR consumers are the CPU mmap path
// (LINEAR de-pad → X2Rgb10 CPU frames) and the VAAPI raw-dmabuf passthrough.
let mut importer = if zerocopy && !backend_is_vaapi && !want_hdr {
let mut importer = if zerocopy && !raw_passthrough && !want_hdr {
if pf_zerocopy::gpu_import_disabled() {
tracing::warn!(
"zero-copy GPU import disabled after repeated import-worker deaths — using CPU path"
@@ -1980,9 +1984,11 @@ mod pipewire {
// host. KWin/gamescope don't need it (they blit into the buffer, so no read-before-render
// race).
let force_shm = std::env::var("PUNKTFUNK_FORCE_SHM").as_deref() == Ok("1");
// VAAPI zero-copy passthrough: zero-copy on, no EGL→CUDA importer (any non-NVIDIA host), and
// the encoder backend is VAAPI → hand the raw dmabuf to the encoder (it imports + GPU-CSCs).
let vaapi_passthrough = zerocopy && !force_shm && importer.is_none() && backend_is_vaapi;
// Raw-dmabuf zero-copy passthrough: zero-copy on, no EGL→CUDA importer, and the frames'
// consumer imports raw dmabufs itself — the VAAPI backend (libva import + GPU CSC) or a
// PyroWave session (the wavelet encoder's own Vulkan device, any vendor) → hand the raw
// dmabuf straight to the encoder.
let vaapi_passthrough = zerocopy && !force_shm && importer.is_none() && raw_passthrough;
// Modifiers our import stack handles for BGRx: the EGL-importable (tiled) set, plus LINEAR
// (0) — NVIDIA's EGL won't list it, but LINEAR dmabufs (gamescope's only offer) import via
// CUDA external memory instead. For the VAAPI passthrough path we advertise LINEAR only:
@@ -1998,8 +2004,9 @@ mod pipewire {
// advertisement with every modifier its device samples from, so compositors that
// never allocate LINEAR (Mutter+NVIDIA) still negotiate zero-copy dmabufs. The modifiers
// were resolved by the facade (`ZeroCopyPolicy::pyrowave_modifiers`) — non-empty only when
// the host's `pyrowave` feature is on AND the encoder pref is `pyrowave` — so capture never
// calls back into `encode` and needs no feature gate of its own (the emptiness check gates it).
// the host's `pyrowave` feature is on AND the session (or the global encoder pref) is
// PyroWave — so capture never calls back into `encode` and needs no feature gate of its
// own (the emptiness check gates it).
if vaapi_passthrough && !policy.pyrowave_modifiers.is_empty() {
for &m in &policy.pyrowave_modifiers {
if !modifiers.contains(&m) {
@@ -2019,11 +2026,11 @@ mod pipewire {
);
} else if zerocopy && !want_dmabuf {
tracing::warn!("zero-copy: no importable dmabuf modifiers — using CPU path");
} else if vaapi_passthrough {
} else if vaapi_passthrough && policy.pyrowave_modifiers.is_empty() {
tracing::info!(
"zero-copy: advertising LINEAR dmabuf for direct VAAPI import (GPU CSC)"
);
} else if want_dmabuf {
} else if want_dmabuf && !vaapi_passthrough {
tracing::info!(
count = modifiers.len(),
sample = ?&modifiers[..modifiers.len().min(6)],